Cloudflare introduced Clef and Clef-flash on October 1 as models for bounded decisions. Instead of treating every classification or routing choice as a long conversation, they score allowed outcomes and return probabilities. Cloudflare offers them through Workers AI and releases the models under Apache 2.0. Cloudflare’s announcement
The distinction matters wherever software must decide among a small set of actions. A model might judge whether a request requires further review, which category it belongs to, or whether a conversation should continue. Restricting the answer to a defined choice makes the output easier to consume and inspect. It does not, by itself, make the decision correct.
Cloudflare reports support for images and a 64,000-token context window. It also publishes a speed comparison across 43 benchmarks: median latency was 209.3 milliseconds for Clef and 38.8 milliseconds for Clef-flash, compared with 524.1 milliseconds for Jev. These are vendor measurements under the stated evaluation, rather than a promise about every application’s response time.
Accuracy comparisons are mixed. On When2Call, the announcement reports 72.37 for Clef and 65.58 for Clef-flash, against Jev’s 80.97. On BRIGHT, both Clef variants also score below Jev. The new models’ lower reported latency therefore should not be read as universal superiority. The faster variant introduces another choice about speed and the quality needed for a particular decision.
My inference is that probability output makes deferral a more explicit product decision. An organization can distinguish a confident prediction from one that deserves human review or a second assessment. But the number beside an outcome still needs to be checked against observed errors in the intended setting. A probability is not evidence that the model is properly calibrated for a new population or task.
Clef’s release gives decision systems a more specialized component and a visible set of tradeoffs. The remaining question is which errors matter and how uncertain cases are handled. Bounded outputs simplify the interface; evidence about accuracy, calibration and review policy determines whether the resulting decision is useful.